# Adaptive Reasoning

> Select one GPT-5.6 Sol worker and calibrate its reasoning effort (low, medium, or high) to task difficulty without parallel execution. Use for non-trivial coding, investigation, review, or planning tasks where a narrow scout, routine worker, or difficult smart worker would improve token efficiency; skip simple questions and one-step edits.

- Skill: `flovencatalasan/adaptive-reasoning` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds@latest add flovencatalasan/adaptive-reasoning`
- Raw SKILL.md: https://api.skillmd.com/api/skills/flovencatalasan/adaptive-reasoning/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Productivity
- Author: flovencatalasan (https://skillmd.com/u/flovencatalasan)
- Updated: 2026-09-22
- Page: https://skillmd.com/skills/flovencatalasan/adaptive-reasoning

---


# Adaptive Reasoning

Match reasoning effort to the work. Use the lowest level likely to complete the task reliably.

## Choose a role

- **Scout — GPT-5.6 Sol Low:** Answer a narrow, read-only question such as locating files, tracing one code path, finding relevant tests, or checking a focused fact. Retain enough judgment to discover useful adjacent context.
- **Worker — GPT-5.6 Sol Medium:** Implement a scoped change, run checks, diagnose a routine defect, or complete supporting work with clear boundaries.
- **Smart worker — GPT-5.6 Sol High:** Handle difficult implementation, cross-cutting reasoning, consequential ambiguity, or a task where a lower effort is likely to miss important constraints.

Handle trivial questions and one-step edits directly without delegation.

## Execute sequentially

1. Classify the next substantive unit of work before delegating.
2. Spawn exactly one worker using `gpt-5.6-sol`, the selected `reasoning_effort`, and `fork_turns: "none"`. Give it a self-contained objective, relevant paths or sources, constraints, success criteria, and required output. Instruct it not to delegate.
3. Wait for that worker to finish before spawning another. Never run workers concurrently.
4. For mixed work, run roles sequentially only when needed—for example, a Low scout followed by a Medium or High worker using the scout's findings.
5. Integrate and verify the result in the primary agent. Keep approvals and consequential decisions with the user.

Escalate reasoning only when task complexity or a concrete failed attempt justifies it. Do not repeat completed work at a higher effort merely for reassurance. If the requested model or effort is unavailable, use the closest available equivalent and state the substitution.